Deterministic agent orchestration · Server-enforced · MCP-native

AIOrc makes your AI agents follow the workflow you define — instead of skipping steps and going off-script.
"One developer builds a flow... everyone uses it."

Your AI agents skip steps and ignore the process you defined. AIOrc fixes that at the root: a server drives execution one step at a time and rejects anything your workflow doesn't allow — so agents can't go off-script or skip a step. Design the process once, share it with your team or the community, and — when you need it — audit every run as signed evidence.

Read the docs →
How it works

Design it once. Run it verified. Prove what happened.

Building agents is the easy part. The hard part is knowing they followed the right process. AIOrc turns a visual workflow into a contract the server enforces — and a record you can audit. No new tool for your team: it works through the AI clients they already use.

Step 01
🎨

Design the flow

Connect agents in the visual Flow Builder. Branch with natural-language conditions on the edges, fork with Parallel nodes, iterate with back-edges.

Step 02
🔌

Connect once

Your team's AI client (Claude Code, Cursor…) connects to AIOrc with a single token. No new app to learn, no per-developer setup — the workflow shows up where they already work.

Step 03
🛡️

Run it verified

The server drives execution one step at a time and rejects any move the flow doesn't allow. Agents can't skip steps or go off-script — the unpredictability that keeps agents out of production, gone by design.

Step 04
📋

Audit & govern

Every run is recorded step by step and exportable as signed, tamper-evident evidence. Pause any project instantly, see which agents actually run, and prove the process was followed.

Capabilities

Everything you need to put agents in production — and keep them in line

Govern, verify and audit your team's agents, and reuse what works across every project.

Two execution modes

Compiled: the whole flow in a single MCP response, zero orchestration latency. Verified: server-driven stepping where illegal transitions are rejected, every dispatch is recorded as ground truth, and any project can be paused or any run cancelled — surgically, without touching the rest.

🔀

Conditional routing

Natural-language conditions on edges, priorities, parallel forks and loops via back-edges with per-agent caps. Decisions and gates are patterns, not extra plumbing.

📋

Agents in markdown

Define each agent as a .md file. The compiler injects its content as instructions when the flow runs.

🧩

Reusable skills

Package specific capabilities as skills and share them across projects and teams. Build on what already works.

🛡️

Fail-closed by design

Structural validation before compiling: single Start, valid agent references, well-formed forks. Cycles are legal loops, capped per agent — and in verified mode the server enforces the caps at runtime.

🔑

Multi-project with auth

API keys per project, user roles, run history and admin panel. Multi-team from day one — not a paid tier.

📚

Contexts: domain apart from capability

Business knowledge lives in Contexts, not inside agents. The same generic crew serves every project — inject the domain, reuse the capability.

📈

Live usage analytics

Which agents, skills and contexts actually run — per project and per teammate, minute by minute. Graph-aware skip detection separates real failures from branches not taken.

Deterministic evals

Test cases per flow, graded by the server against the verified execution path. Know whether a change to an agent improved the flow or broke it — before your team finds out.

🗂️

Signed audit trails

Every run is reconstructable step by step: who ran it, which transitions it took and why. Export any run as HMAC-signed JSON — tamper-evident evidence, aligned with EU AI Act logging requirements.

🔗

One connection to rule them all

Your client connects to AIOrc once. Workflow tools and any external MCP servers you register are exposed through that single connection — proxied, namespaced and logged.

Architecture

The compiler, under the hood

compileFlow() walks the graph, emits an instruction block per node and concatenates them into an ordered payload for compiled mode. In verified mode the engine goes further: it keeps the state machine on the server, dispatches one step at a time and validates every transition against the graph's edges.

🖼️
Canvas
Flow Builder
Start, agent, parallel and end nodes connected visually
🔁
Compiler
compileFlow()
Topo-sort · recursive walk · convergence · validation
📜
Output
Instructions
Context + skills + agent blocks + topology & execution rules
🤖
Execution
Target model
Claude executes each step; in verified mode the server confirms every transition
// Compiled instructions — sample fragment FOLLOW the topology. After each agent, take exactly ONE of its allowed transitions — never skip or invent steps. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ## Agent 1: task-planner Analyze the requirement and produce an action plan. **Expected output format:** Structured plan with tasks ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ## Transitions from task-planner → security-qa when: the plan touches auth or tenant data → code-reviewer (no condition — fallback) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ## Agent 3: code-reviewer Review the implementation. Score 0–10. { verdict: 'PASS|FAIL', score: N }
The platform

Designed to work, not to configure

From the visual canvas to the project panel — all in one interface built for anyone building with AI agents.

Flow Builder — agent canvas

Flow Builder

Design multi-agent flows visually. Connect nodes, set edge conditions and parallel branches — the compiler does the rest.

Project panel

Project panel

Manage all your projects from one place. Agents, skills and flows organized and ready to share with your team.

Who AIOrc is for

Built for anyone shipping AI agents — solo or in a team

👥

Collaborating teams

Share flows and agents with your teammates. Every team member builds on what others have already tested, without starting from scratch on each project.

🧩

Skills as team assets

Package specific capabilities as reusable skills. A well-defined skill becomes available to all projects and developers on the team.

🌐

Flow community

Explore flows and skills published by other users in the community repository. Adapt what already works and contribute your own.

🏢

Multiple projects, same know-how

Keep a library of specialized agents and deploy it across different projects. Team knowledge accumulates, not repeats.

🏗️

Development pipelines

Planner → Developer → Code Reviewer → Security QA. Deterministic CI flows any team member can run with a single call.

📝

Standardized processes

Define the process once, execute it consistently. From code review to content generation, with the same agents and the same rules every time.

🔍

Engineering leads

See which of your agents the team actually relies on and which are noise. Audit any run step by step, with every transition validated and recorded by the server.

Origin

Context

As a Software Engineer and AI researcher at my company, I was assigned to evaluate the state of the art in LLMs and agentic models. My recommendation was to adopt Claude Code. That's where the real problem started.

Two paths to orchestrate agents

I first tried to orchestrate agents deterministically using only the models. It didn't work: LLMs are probabilistic and often choose the path of least resistance instead of the correct one. The model ended up creating its own agents and skipping the ones I had defined.

Then I built a plugin for Claude with a main orchestrator, 21 subagents and 17 skills, using agent teams. It worked better, but the feature is still experimental and the bugs made it too unreliable for team use.

That's when I asked myself: what would a layer that runs deterministic prompt flows with agent orchestration, geared toward developers and dev teams, look like today? That's how AIOrc was born.

The problem it solves

At my company we have multiple projects with different stacks: a Nest.js backend with TDD, blockchain integration, fund custodians, AWS, MongoDB Atlas. A frontend with design patterns polished over years. If I wanted to use the flows my coworker had built, I had to ask them for agents, skills, context, copy them into my repo and hope Claude would follow the flow.

With AIOrc, my Claude connects to an MCP with a token. No cloning, no copying, no drift. When I migrate to Python tomorrow to add data science, I go to the public flow repository, grab what I need and use it.

A cross-disciplinary layer

We want AI to be universal. If we give Claude to a graphic designer, or if we give a backend developer who doesn't know styles the frontend flows from the team, the logic is exchanged. A developer shares the flow and the entire company — or the community — can follow their standards.

Where it stands today

Everything is oriented toward agents and automation with a clear goal: reduce execution time, gain reliability in the LLM and eliminate slop — that scattered, unreliable output models generate when they don't have a deterministic flow to follow.

It's just getting started. Today you can already create projects, orchestrate flows in compiled or server-verified mode, define eval suites that grade runs against the graph, watch live usage analytics per agent, skill and teammate, share everything with your team privately, make public what you want and consume others'. The idea is to build a community of experts in web2, web3, blockchain, Python, C# and everything else we can think of, sharing secure databases, cloud connections, interface design and software architecture.

AIOrc is still a prototype in its early steps. What it needs now are people who try it, break it and share what happened. Early feedback is what defines where it goes, and the community is what will turn it into something truly useful. If what I'm building resonates with you, join in.

Ready to start?

Start orchestrating today

Create your account, design your first flow and share it with your team in minutes. Multi-team workspaces, permissions and audit — the features others put behind enterprise plans — are simply included.

Read the docs